Open coding models

Can anyone run a top coding model themselves?

Contributes to: Can AI do the work?. Signal score: 67 / 100. Relative weight within this question: 1.

StatusFalling behind
Latest value4 months
Caution

This only uses models Epoch AI has tested on SWE-bench Verified, and only those models.dev lists, so a new model can be missing for a while.

Open-weight means anyone can download the model, not that it is cheap to run. The largest open models need serious hardware.

Observed trend

Months behind closed models, at each new best open-weight score, by release month.

Months behind closed models, at each new best open-weight score, by release month. · 2026-01 – 2026-06

2026-06 — Months behind closed models, at each new best open-weight score, by release month.: 4
View data table
Period / observationMonths behind closed models, at each new best open-weight score, by release month.
2026-012
2026-042
2026-064

What it measures

Metric
Months the best open-weight model trails closed models on SWE-bench Verified
Solved if
Open-weight models match the best closed models as they are released
Signal score
67 / 100. The best open-weight model trails closed models by 4 months, against a goal of none. A year behind scores 0%. How it is scored
Status rule
Catching up when the newest open-weight best trails closed models by fewer months than the first one did.

Each new best open-weight score

The best closed model scores 83%, from claude-opus-4-7, against 79% for the best open-weight model.

Update history

The reading on each date we checked, separate from the historical observations above.

DateValueYear over yearStatus
2026-11-074 months—Falling behind
2026-11-064 months—Falling behind

Source

SourceRoleMeasuresLatest data
models.devScored measurementWhich AI models have open weights2026-11